Research On Segmentation Algorithm Of Prostate In MR Image Based On Deep Learning

Prostate cancer is the second common cancer among men.MRI is the main method for observing the lesions of the prostate area because of its clear imaging of human soft tissue structure.However,the imaging area of the MRI is the entire pelvis of the human body.The small size of the prostate and the close proximity to the surrounding tissues bring inconvenience to the clinician’s observation.Manual segmentation of the prostate requires a lot of time.Therefore,the automatic segmentation of the prostate from MRI is of great significance for the clinical diagnosis of the disease.The convolutional neural network with U-shaped structure as a framework has been very outstanding in image segmentation tasks in recent years,and has almost become the standard framework for image semantic segmentation in deep learning methods.The MRI of prostate is medical image,which has the characteristics of low contrast,small target volume,unclear boundaries,and few available samples.It is also difficult to segment compared to natural images.Therefore,the basic U-Net can’t accurately segment the prostate.According to the characteristics of prostate MRI,a set of segmentation algorithms is proposed: first,the image pre-processing module normalizes the image size and gray value,and enhances the image contrast by adjusting the window width and window level and histogram equalization.The curvature filter smoothes the image;then the training samples are expanded by non-rigid body transformation,affine transformation,horizontal flipping,random cropping;finally,based on the U-shaped network architecture,the ordinary convolution layer is replaced in the encoder with dense connection structure to enhance the ability of the neural network to extract features while reducing the amount of parameters.In the decoder,an attention mechanism is added to learn key features.In the prediction output,contour prediction branche is added to enhance the ability to learn prostate contours.Based on the published prostate datasets Promise12 and Prostate X,the proposed segmentation algorithm is experimentally verified,and its DSC and Hasdorff distance can reach 0.89 and 10.1,respectively,and it has fewer parameters than other network structures.The experimental results show that the proposed segmentation algorithm can effectively segment the prostate,and it has the advantages of high segmentation accuracy,few parameters,and fast model convergence.